Advancing multi-domain active learning: strategies and techniques using neural networks for classiﬁcation

He, Rui ORCID: 0000-0002-6338-1387 (2024). Advancing multi-domain active learning: strategies and techniques using neural networks for classiﬁcation. University of Birmingham. Ph.D.

Preview

He2024PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (4MB) | Preview

Abstract

Dealing with data collected from various domains has emerged as a pervasive challenge in machine learning applications, necessitating the use of multi-domain learning (MDL) to efficiently process and learn from the collected multi-domain data, which have different distributions. Nevertheless, constructing high-quality annotated multi-domain datasets can be both challenging and expensive due to the difficulties in accessing multiple domain experts. Active learning (AL) provides a potential solution to this issue, effectively reducing labeling costs by only selecting and acquiring labels for the most informative data instances. However, conventional AL methods are not designed to handle multi-domain data and cannot be directly applied to multiple domains. Moreover, there is a notable scarcity of studies exploring the intersection of MDL and AL. To fill this gap, the focus of this thesis is on multi-domain active learning (MDAL), an innovative approach that combines active learning and multi-domain learning, enabling cost-effective learning through the use of neural networks.

In this thesis, we first conduct a comprehensive review of AL and MDL research, formalizing the problem definitions and settings for MDAL. An open-source awesome active learning knowledge base is also developed to facilitate our research and the community. Second, to address the problem of MDAL, a unified pipeline is proposed to integrate conventional AL methods with neural network-based multi-domain learning models. Extensive comparative experiments reveal that certain model-strategy pairs yield strong performance empirically across domains. Third, to address model limitations under insufficient annotations, the thesis introduces a plug-and-play multi-domain contrastive learning method applicable to various MDL models under share-private architecture. Experimental results demonstrate significant improvement in model performance with insufficient annotations. Fourth, limitations of conventional AL methods in assessing domain-shared information motivate the development of tailored MDAL algorithms. From the labeling strategy perspective, a novel perturbation-based two-stage AL strategy is proposed to select informative cross-domain instances. This approach outperforms conventional methods by explicitly evaluating domain-shared information.

Through formalizing the problem, extensive comparative analysis, and introducing new learning methods and strategies, this thesis advances the state-of-the-art in cost-effective MDAL using deep neural networks. The proposed techniques and findings provide valuable insights to facilitate future multi-domain learning applications with limited labeled data.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):